Dominating an Online Game with Object Detection Using OpenCV - Template Matching.
HTML-код
- Опубликовано: 11 май 2024
- Detect objects with No GPU, No Neural Network, and No training. Template matching has some unique advantages including being really easy to set up. You only need a few lines of code to get started. You can also do it without any high-end computing resources.
In this video, I walk you through a simple example of how to use template matching and then show you how I used it to dominate an online game.
github.com/ClarityCoders/Comp...
Want to chat with me and other programmers join our discord!
/ discord
Want to help support the channel?
www.buymeacoffee.com/clarityc... - Наука
Great video, love to see more like this. Automating games like this is an amazing way to introduce people to OpenCV, it's simple to explain but teaches you important skills applicable to a much wider range of tasks. I'm glad I randomly stumbled upon your channel, your videos don't get nearly as many views as it deserves for such high quality educational content.
Glad you enjoyed it! I'll have more coming soon!
I agree, I had my first coding experience from minecraft because I wanted to automate a turtle in the modded game xD
@@yichong777 Thanks for the comment it means a lot. I was getting beat up a bit today haha...
i bascily became a programmer by starting to atuomate games. learned many skills from basic programming to ocr to stock market theroies to reinforcement learning while doing that.
@@bobchelios9961 So did i....... which is why it's kind of a shame some games are so quick to ban.
Amazing work! Thank you so much for sharing this! I would love to see a video on how to implement convolutional neural networks to do object recognition.
It's on my list for sure!
This is a really well paced and detailed guide! Great work :)
Thanks I work really hard on trying to figure out the best pace to keep people engaged and still not skip topics. I will continue to try and improve. Thanks for commenting and watching.
Your videos are awesome man! I really love them :D
Mean a lot thanks. I really appreciate the view and comment have a good one.
Great tutorial, really enjoy and appreciate you explain what each line does in detail. Subscribed and liked!
This is an awesome, very straightforward tutorial. Thank you!
To circumvent the missed image detection from a day/night cycle, you can add a desaturation to both your sample and comparison images to grayscale them before comparison.
This usually works for pixel-based art styles, as day/night is usually just an added color filter over the original image.
The exception being those games that use specular or bloom lighting effects, which can distort the end-result grayscale in the image.
The trade off, however, is that the added step of desaturating the sample image on each frame will slow down the script a bit.
EDIT: This method also wouldn't work in games that reuse sprites with different color palettes, for obvious reasons.
Yes sir! Very good points this is something I've used in a few projects and probably should have talked about! Great comment.
hey could you help make a open cv for shooting the ball in nba 2k
I didn’t get it why do they use different color palettes for different reasons?
Differentiate objects by color is a cost effective way of expanding the object types in game.
Red goblins are fire goblins.
Blue goblins are ice goblins.
If you want to use flame thrower on red goblins, and water gun on blue goblins, you need to be able to determine type you're identifying.
That's the issue discussed above.
Ah thank you so much, i was looking for a tutorial like this for a very long time :) I did something similiar in auto it just for fun, but I always wanted do something like that in a "real" programming language with more powerful libraries. Thank you so much =)
I only need to learn how to use it in 3d games /video capturing. But it is just fine to learn right from the start :)
Thank for the examples and how it all works, liked and subbed.
3:34 "now my hope is that we detect multiple instances of weed"
yes pls!
except he clearly says wheat
@@NoshNosher are you sure about that bro?
@@NoshNosher Charles Shwab ova here
@@Beatsbasteln I am 100% sure he said wheat.
@@emifro but how are you going to smoke wheat?
I love to see you doing these videos! I’d like to see one that is based on real pictures, like “is the car pointed left or right” etc and incorporate some training and AI
Cool idea! I'll put it on my list.
Thank you man, this is awesome !!
Glad you liked it thanks for watching!
haha this is awesome, i've thought about doing it before and after this i might have to. One of the most relatable/interesting coding videos lmao.
glad you liked it! I hope you do please share it if you do.
@@ClarityCoders I already did :) although apparently the only way to grow small channels these days is to leave comments on a bunch of related videos so people click on your profile.
@@stevesteverson3345 Trying to grow your channel?
@@ClarityCoders hahah yeah come watch any of my 0 videos
Liked and subscribed - explained and shown structured and clearly! Thanks!
Oh, come on.. I just decided "Ah crap, just get started..(!) Ima start with the basic webstack and see from there.
And then I stumbled over this channel and THIS video..
Great video @ClarityCoders!
thanks man. I love your teaching style
I love the tutorial; keep up the great work! Heads up, I am coming for your high score ;)
I hope you do! Thanks for watching I appreciate it!
Your score has been topped! No template matching was needed since the game uses the exact same pixels in the same locations but still great to know that functionality exists within OpenCV. keep on keepin on.
@@wyatt5984 Yeah I noticed that as well haha! Gives you some flexibility for more complex games though. Nice score!
@@wyatt5984 Our discord is you're into that sort of thing. discord.gg/cAWW5qq
You could have also done the grouping by comparing the distances of all rectangle centroids which would essentially create a node graph with every center connected to everyother center and then combining all nodes whose distance is below some threshold distance. Your solution works, I just think it’s good practice to use more general algorithms.
I would agree very good points. Thanks for watching and commenting.
Thank you for making this video, it really helps me a lot.
I hope so. Let me know if you have any questions.
Ty very much, i could not find someone explain this as well as you did ngl saved me
this is what I have been looking for... THANK YOUUUU
Teaching archivement unlocked! congrats man! you are helping people from brazil!
I am from a tiny town in the US never thought people would watch me in Brazil. Thanks for watching and commenting.
@@ClarityCoders 🇧🇷
Very fun to watch and learned a lot. Thankyou.
Thanks for watching/commenting it means a lot.
15:15 A more efficient solution here would be to just check for rectangles with intersecting regions. Not sure about python but in c# this is very simple with the Rect.IntersectsWith method (you can also just do it manually, the math is easy). Then you can combine them by inflating one rect with the difference from the other.
Also for more reliable matching, there are various imaging algorithms you can use, for my OCR projects I do: grayscale > scaling > binarization > inversion > denoise > bordering, this helps a lot for matching foreground objects over varying backgrounds. You can also start with colour averaging which can make scanning for objects significantly faster at higher resolutions because you can scan chunks of the image and skip processing any that aren't similar. I've used these methods for overlaying real-time translation of text in non-english games.
Nice; Solid introductory video! I enjoyed watching. Wish there was a fast & efficient template matching method that worked with a high dimensional image space. I imagine the game you tested in the vid was only 800x600 or something lol. When working with large images I usually have to search a small section or it will take too long to compute.
For sure I almost always cut down the area where I'm taking the screen shots. Especially if it's a real time thing where time matters. Thanks for watching and commenting I really appreciate it.
You could also try speeding it up by scaling the image down before performing the template matching.
@@SuperMasterDesaster Yeah I really like this idea as well.
Couldn't be more clear. Many thanks...
best tuto about template match i have ever seen!helpful!
Thanks man, you saved my graduation project!
Wow, this is so cool! Thanks for sharing!
Thanks for the comment more to come stay tuned.
Dude, this is so genius. Thx!
amazing. thanks. so much easy class from you
Pretty clear and very interesting, coming from C and learning Python and opencv for industrial purpose, you helped me a lot , template matching can be used in so many way in my work. you have all my respect and i'll follow you with great joy. Thanks again for your videos.
This comment really means a lot to me. Thanks for watching and commenting I'll keep making videos.
Awesome video! Learned a lot!
I'm loving your content! As a fellow ML engineer, and fan on mini online games, I love this!
Thanks! I really appreciate the feedback and views!
Pretty nice! Found this on my recommended. Subbed :)
Thanks for the sub / comment means a lot.
Very useful, thank you for sharing!
Thanks for watching and commenting!
Thank You for this valuable video
great job man :D
Thanks I really appreciate it.
woah you jump from 2k subs to 7k in 5 days, congrats
Great video, very helpful
Here's an idea to get around the different shades thing: transform all target images to black and white, and same with the recording output going into the program.
Different colours will end up as different shades of grey though. The correct approach is what the library already does, which is normalisation.
@@gnack420 Shades imply there's color. What you wanted to say was value (lightdark). Sorry I got into definitions. But yeah, you're right. I wrote that comment before finishing the video, I didn't know you were gonna use an algorithm to find likely matches, I thought that whould require some deep learning or something much more advanced. Shows how much I know about programming. Oh, and btw, thanks for the video, it was really fun!
Great video thank you!
This is a huge inspiration.
thanks means a lot!
this seems really really cool; I want to do that on forge of empires . I'll watch this at some point
Hope you do! Thanks for the comment.
Great tuto!
Amazing video!
ty cuz ive been having a hard ti getting started.
Anytime let me know if you have any questions!
blessed by youtube algo.. found your video and subscribed!
It's actually given me some love this past week. Thanks for watching and commenting.
In The setup thank you so much!
No problem. Thanks for watching.
Dame! Great job!
Really cool!
This is really cool!
Thanks for watching and commenting.
👏 Thank for this
Love this
Thanks the comments and views really mean a lot!
Great Video!
Thanks for watching protect that melon.
This is a great tutorial - thanks! Could you also do one explaining the kick ya chop script you have in your git hub? Also I tried to use the kick ya chop script but had to increase the sleep time to around 0.2 otherwise it wouldn't work. This is even after I tried reducing the region to search in. What gives? Do I need a faster computer or something?
Yeah it might just be a difference in processing speed. I'll try to get some videos out explaining the code more in depth.
OMG, thank you!!!!
Amazing tutorial although I do have some ideas to improve some of what was presented,
Firstly the rectangles grouping section, I'm sure there's a better way than just duplicating the list, maybe manually comparing the x, y of rectangles and if they are similar to another rectangle we don't use it
The other point would be about beating the highscore, i would test for a single pixel's color above the player on each side where the background is static and check if it's color changes (a branch is above) and move accordingly
Very good video!
Thanks for watching Vinicius and for the comment. Means a lot!
For the multiple boxes issue, could you lower the threshold till at least 2 boxes appear around the wheat, then remove any 1 box items which should be anything that isn't wheat, and continue from there?
This is awesome thanks for the tutorial!
I just wonder how to implement the code with 5 multiple window (open at the same time)
Nice video but one question. Did you purposefully excactly double their highscore at the end or was that coincidence?
I didn't even know! Good to have eyes like yours on these videos. Thanks for watching Noah.
Here after watching CS Dojo's video :)
Awesome thanks for stopping by!!
Thank you.
No problem!
ClarityCoders
-very very awesome
Thanks I really appreciate the comment.
@@ClarityCoders DO MORE AWESOME THINGSS)))
thanks learned a lot :)
Thanks! Hope you stick around more cool things coming.
Amezing tutorial, subscrubed!
Thank you! Thanks for watching / commenting.
Very good video, can you combine your idea into the online game of chess, and use OPENCV to identify the FEN position code of chess on the web page?
You can also find the wheat points without overlaps by running a local maxima filter on the grayscale image- accomplished in OpenCV with a dilate filter followed by a compare filter. You might not need it for this use-case, but you can also additionally filter out plateaus of local maximum value with an erode and a compare (if you want to know what to compare against what, for the sake of brevity here either google it or just try a few combinations; heck, you don’t need Jupyter to figure out the correct operations; you can prototype this kind of stuff with Photoshop/Gimp layers easily).
I like image-processing solutions like this because 1. For an end-use you can more easily run the whole thing on a GPU where operations over a whole image are cheap and reducing that an array of numerical results is the expensive part, and 2. You can visually see and tune/debug the dilated image, the dilated-compared image, the dilated-compared-eroded image, and the dilated-compared-eroded-compared image.
Very good points and great feedback. Thanks for watching and commenting.
Legends say the guy is still punching.
great tut. btw, what is the screen capture software used to make the vid?
OBS
@claritycoders Thanks fo the videos! May you guys comment on how do games (their developers) catch such bots? Asking for the ways around obviously)
Man that's awesome
Thanks! I really appreciate the comment and watching.
Impressive!!!
Thanks for watching I really appreciate it!
2nd year cs student - these videos are awesome keep it up!!
Hi brother is it possible for you to make a chess bot with opencv as u make for other games or for knife hit game if yes please make a video on it .It will be really helpful .
Sure let me see what I can do! Hit that bell so you don't miss it.
ohh I already done that as I like your content so much and thanks it really means a lot.
@@grzegorzabedzki2989 awesome work, very readable code, even though I have no interest in chess.
@@ClarityCoders please don't, chess sucks with cheaters. And if you must, at least don't release the code
you are wonderful
thanks🤩
Thank you! Very great guide to OpenCV! But how to optimize it to get it to run 100 times faster? How to use multiple CPU cores and GPU with OpenCV?
@TridentflayerThe other thing you can attempt to do is vectorize your code very heavily so that you can do all of your calculations in numpy operations. Speaking from miserable experience, that can be far harder than just biting the bullet and doing it in c++
very interesting, subscribed :)
Thank you sir!
This is AWESOME and it has been extremely helpful to me.
There is one thing I'm trying to figure out. Let's say I grab an image and I know I want to look for one "needle" in the lower half of the image and a second in the upper half of the image and I want to know the relationship of the matches to each other.
Now one way I could do this is grab two different regions two images. Do the matches, and then use some math to figure out their relationship to each other.
But, what would be really convenient is if I can just look for the matches I need to in the lower half the image and get, multiple matches and then look for the other needle in the upper half and get multiple matches and already have the x and y coordinates for the overall image without having to translate them.
If it can't be done, it's easy enough to add y offset to the coordinates I get out. But I was wondering.
You mean you want to save time by only looking in the correct regions of the images while searching? You could use slicing to accomplish this task. Google python slicing.
great video man, terima kasih
Thank you sir.
Teacher, first congratulations for the channel. I need your help, I have a simple image with a range of 5 placements where the quantity is identified, how do I identify the biggest color type red and say where the X,Y is, and if you can help me, and only for create an alert.
expected to see dominating that online farming game
Sorrry! Maybe in the next one 😏
thank you
You're welcome!
Holy shit I need to do this.
Jesus!!!!! Love it
Thank you Mac Miller.
Hahaha love it.
💯 th like
💯+💯th Like
TNice tutorials video is a great introduction to soft soft for beginners. I would Nice tutorialghly recomnd tNice tutorials tutorial to anyone starting out with soft soft.
Appreciate it! Thanks for the comment.
Jupyter us perfect for a live demonstration 👌
I agree. I actually caught a lot of heat for using Jupyter haha...
getting game information from memory: 😴😴🤢🤬🤬
getting game information using opencv: 😎😎🤑🤑🥶🥶
Both can be crazy frustrating haha...
Oh, this might solve a problem I have. There’s some made up alphabet which looks like it was made with a fixed typeface, and I want to transcribe the text in the images into actual text in order to try to decipher it, and I think I know how to extract the characters, but I still need to find how to detect which character they are.
I can’t use standard OCR because these are a made up alphabet in an Easter egg for a game.
that small insect walking across his screen got me
Dang bugs.
Are you taking screenshots from the whole screen or just the browser content?
Maybe I could use some of this knowledge to augment Serpent into playing something like hades or issac again.
serpent?
@@YuikyuWolf SerpantAI
Hi i am beginner, i am preparing script for a similar game but the question is can i run script like this for multiple game windows lets say arranged in 3x3 matirx (just like tic tac toe) on desktop and would it work or do i need to implement something more?
great video! could you make something like this for a pokemon game walking in a grass and leveling up?
I always compared images with mean square root or structural similarity. Template matching seem to be more intuitive
Probably similar in speed? Thanks for watching / commenting.
Is it possible to move the mouse to the location of the identified object on the screen instead of a pre-set point?